Yes, but I'd say there is also responsibility attached to the deployment and use (be that, e.g., for quant fund performance, behaviour on exchanges, ...). In a way, using an algorithm is the decision humans make that matters.
At the same time, I've been involved in hundreds of After Action Reviews of incidents and it's not always simple or easy to both figure out which party was ultimately responsible and also how to allocate the "dollar error budget" to the parties overall.
Why would that need a new way? Or do you mean that no-one ever spun up anything at all (e.g., agent zero was never spun up or allowed to be spun up by anyone)?
That's the billion dollar question. Nvidia currently wants to deploy chips with the sole purpose of monitoring agentic workloads and that doesn't seem farfetched but they obviously have a financial motive to sell more products. It's like a cat and mouse game, like cybersecurity in general.
It isn't some law of nature. Humans/societies have agency - what AI should or should not be used is up for debate and decisions. It might even wind up the other way around that using AI is the special case - who knows.
> what AI should or should not be used is up for debate and decisions.
Of course; but it's very hypocritical to raise these feelings only when mathematicians are affected, whereas all the above professions are just told to adapt to the new way of things.
For sure though, translators don't have the same clout and social status as mathematicians do.
In some form and shape, yes. Humanity's creativity is a lot of marginal copying and remixing.
But obviously, it adds up to something greater than went in; in aggregate, our contributions are something to awe.
But my point is, if you zoom in at the marginal, incremental contributions of any individual human in this process, it's really hard for me to say LLMs are not at the same level already.
On this topic, people like to compare LLMs to Einstein, but as far as I know, Einstein did not zero-shot special relativity in an afternoon. He built it up incrementally over time, it took him three times longer than the time between first ChatGPT release and today, and it depended on centuries of prior art, culminating in the right observation and right notation being available to him in his moment of greatness.
What would you accept as evidence there? Are, for example, the first names/words for colors from scratch?
So your view is that everything was there at the creation of the universe (it's a possible view, of course)? Or are there any "things" that can create ideas from scratch?
Recently I watched a documentary on the tanzanian Hadza tribe, one of the last hunter gatherer tribes on Earth. Their language is a distinct click and pop language and they regularly imitate animal calls (monkeys, baboons, birds) when they hunt but also when they communicate with each other, tell stories etc.
I think it's not impossible that words evolved as adaptations of the environmental sounds with which our ancestors lived. The human creativity producing DNA is also a remix of preexisting molecules formed under evolutionary pressure, so the view that it's turtles all the way down, unintuitive as it is, may not be so indefensible after all.
I mean what is your criterion on invention here? On the one hand, each specific word could be seen as a new invention. On the other hand, all languages basically correlate strongly with the environment of their users - it's why LLMs turn out to be universal translators - and pattern-matching is hardly an invention, isn't it?
I don't know, but my current intuition is that invention = more or less preexisting ingredients + some tiny spark of creativity ("connecting the dots") rather than this grand myth of "creating something from nothing". And then the tiny spark could itself be a remix of previous tiny sparks (connecting the dots between dot connections, a bit like maps between n, n+1, n+2, ...-dimensional manifolds), and so on.
> and pattern-matching is hardly an invention
Exactly. That's what I meant with turtles all the way down, until the tiny spark disappears into a dot and it turns out that creativity is just an illusion - akin to how nobody would call the gene (the spark, the pattern matching) formed by evolution (the brain connectome) "creative".
My view is that LLMs meet the standards by which we judge human creativity/inventiveness, and thus that one cannot claim LLMs "just repeat, never invent" without the same being true about humans.
At least in Sid's case, it went from the oncologist saying "I have no more drugs I would recommend, no trials available" (slide 7) to "I currently have no evidence of disease" (slide 18). I don't know beyond that or beyond Sid's case - or a similar story of an Australian who treated a cancer tumour their dog had with a similar AI / personalized vaccine process.
My understanding of what Sid's describing is that you do RNA sequencing, a whole genome sequencing, feed that into frontier AI (if it will still let you), and somewhere along the way give the information the AI finds to people who can use it make a personalized mRNA vaccine, specifically for you and your cancer.
Another link here about Sid's case, it explains it didn't go through trials: "made possible through a compassionate use allowance from the U.S. Food and Drug Administration (FDA)".
Hand counting paper ballots by a somewhat random assembly of people might make that less easy (as people might need to collude). Also, why would the ledgers be accessible to everyone counting? Other than using them to check off who voted/eligibility when a voter shows up, what are they needed for at the counting stage?
The decision to publish the ledger is one made by the election administrator in the interest of transparency - there is no requirement to do so, but it is believed to increase public trust in the process. Unfortunately these types of transparency measures haven't necessarily been rigorously evaluated and the tradeoffs with ballot secrecy may not be favorable. that is the real significance of this research to me, not really anything about the use of machines.
Some states do quite the opposite, e.g. seal ballots after the election and only allow inspection by court order. This has obvious benefits for secrecy, but then creates a lot of objections when people/organizations that doubt the election outcome are told that they cannot review the actual ballots (without meeting some threshold to justify such an order, which can be quite high since these rules were often put in place to try to counter pay-to-vote or voter intimidation schemes).
The existence of such a ledger isn't even required, but it makes it radically easier to audit the count process to detect errors. US election administrators demand a very high level of accuracy in vote counts (higher than what is typical in a lot of other countries), so in hand-count processes it's common to tabulate each ballot multiple times, and with machine tabulation there is usually some kind of automatic recount or risk limiting audit, either by retabulating with separate machines or hand-counting a sample. All of these are much easier processes when you can correlate a ballot to the previous time it was counted, so that errors discovered in auditing can be tracked to the specific ballot that was miscounted and corrected. Otherwise, if you count a batch of ballots twice and get different results (which is virtually guaranteed with hand counting), you will have to recount the entire batch over and over until you get satisfactory convergence. There are several different methods of doing this: some jurisdictions serialize ballots when they are printed so that tabulation can be tracked by that serial number. Other jurisdictions avoid this potential secrecy hazard by only serializing ballots after initial tabulation (e.g. ballots pulled for a sample audit have sequence numbers added on stickers or hand-written). Some tabulating machines can add sequence numbers as they run. All of these have different considerations as far as reliability and secrecy, but unfortunately there is next to zero funding or political interest in researching these issues and applying the resulting knowledge. The result is that everything is highly variable from state to state and some states continue to use processes that are known to have significant defects, most famously Louisiana with its ongoing use of non-auditable DRE machines (the only state that continues to do so).
In this case, the ledgers were published on the internet, there was no collusion necessary to obtain them. The problem here is more the level of record-keeping and transparency about the counting process, not how the counting itself was done.
reply